new optimization of dgemm kernel for bulldozer: 10% performance increase

optimized again bulldozer dgemm kernel

Develop dgemm for bullozer

Instead of using thread 0 buffer, each thread uses its own sb buffer.
Thus, it can avoid overflowing thread 0 buffer.

Standard headers may contain C++ templates which are not permitted inside an
extern "C" block. This might be the case when we include <complex.h>.

Thank Dan Povey's patch. https://github.com/xianyi/OpenBLAS/issues/187

The 'const' modifications were done automatically using this scripts:
https://kaldi.svn.sourceforge.net/svnroot/kaldi/sandbox/dan/tools/for_openblas

Netlib style:
    enum CBLAS_XYZ {X=1, Y=2, Z=3};

Intel MKL style:
    typedef enum {X=1, Y=2, Z=3} CBLAS_XYZ;

With this hybrid style, code written in the latter form won't need any
modifications to be built with OpenBLAS.  This change should not affect existing
code, although a warning may be emitted for C code which does the following
(does not occur with C++):
    typedef enum CBLAS_XYZ CBLAS_XYZ;
    warning: redefinition of typedef 'CBLAS_XYZ' [-pedantic]

set parameters for CORE_ATHLON

else dgemm_p is set to zero leading to a segfault in alloc_mmap due to
allocsize being zero